networking/zcip.c

   1 /* vi: set sw=4 ts=4: */
   2 /*
   3  * RFC3927 ZeroConf IPv4 Link-Local addressing
   4  * (see <http://www.zeroconf.org/>)
   5  *
   6  * Copyright (C) 2003 by Arthur van Hoff (avh@strangeberry.com)
   7  * Copyright (C) 2004 by David Brownell
   8  *
   9  * Licensed under GPLv2 or later, see file LICENSE in this source tree.
  10  */
  11
  12 /*
  13  * ZCIP just manages the 169.254.*.* addresses.  That network is not
  14  * routed at the IP level, though various proxies or bridges can
  15  * certainly be used.  Its naming is built over multicast DNS.
  16  */
  17
  18 //#define DEBUG
  19
  20 // TODO:
  21 // - more real-world usage/testing, especially daemon mode
  22 // - kernel packet filters to reduce scheduling noise
  23 // - avoid silent script failures, especially under load...
  24 // - link status monitoring (restart on link-up; stop on link-down)
  25
  26 //usage:#define zcip_trivial_usage
  27 //usage:       "[OPTIONS] IFACE SCRIPT"
  28 //usage:#define zcip_full_usage "\n\n"
  29 //usage:       "Manage a ZeroConf IPv4 link-local address\n"
  30 //usage:     "\nOptions:"
  31 //usage:     "\n        -f              Run in foreground"
  32 //usage:     "\n        -q              Quit after obtaining address"
  33 //usage:     "\n        -r 169.254.x.x  Request this address first"
  34 //usage:     "\n        -v              Verbose"
  35 //usage:     "\n"
  36 //usage:     "\nWith no -q, runs continuously monitoring for ARP conflicts,"
  37 //usage:     "\nexits only on I/O errors (link down etc)"
  38
  39 #include <netinet/ether.h>
  40 #include <net/ethernet.h>
  41 #include <net/if.h>
  42 #include <net/if_arp.h>
  43 #include <linux/if_packet.h>
  44 #include <linux/sockios.h>
  45
  46 #include "libbb.h"
  47 #include <syslog.h>
  48
  49 /* We don't need more than 32 bits of the counter */
  50 #define MONOTONIC_US() ((unsigned)monotonic_us())
  51
  52 struct arp_packet {
  53         struct ether_header eth;
  54         struct ether_arp arp;
  55 } PACKED;
  56
  57 enum {
  58 /* 169.254.0.0 */
  59         LINKLOCAL_ADDR = 0xa9fe0000,
  60
  61 /* protocol timeout parameters, specified in seconds */
  62         PROBE_WAIT = 1,
  63         PROBE_MIN = 1,
  64         PROBE_MAX = 2,
  65         PROBE_NUM = 3,
  66         MAX_CONFLICTS = 10,
  67         RATE_LIMIT_INTERVAL = 60,
  68         ANNOUNCE_WAIT = 2,
  69         ANNOUNCE_NUM = 2,
  70         ANNOUNCE_INTERVAL = 2,
  71         DEFEND_INTERVAL = 10
  72 };
  73
  74 /* States during the configuration process. */
  75 enum {
  76         PROBE = 0,
  77         RATE_LIMIT_PROBE,
  78         ANNOUNCE,
  79         MONITOR,
  80         DEFEND
  81 };
  82
  83 #define VDBG(...) do { } while (0)
  84
  85
  86 enum {
  87         sock_fd = 3
  88 };
  89
  90 struct globals {
  91         struct sockaddr saddr;
  92         struct ether_addr eth_addr;
  93 } FIX_ALIASING;
  94 #define G (*(struct globals*)&bb_common_bufsiz1)
  95 #define saddr    (G.saddr   )
  96 #define eth_addr (G.eth_addr)
  97
  98
  99 /**
 100  * Pick a random link local IP address on 169.254/16, except that
 101  * the first and last 256 addresses are reserved.
 102  */
 103 static uint32_t pick(void)
 104 {
 105         unsigned tmp;
 106
 107         do {
 108                 tmp = rand() & IN_CLASSB_HOST;
 109         } while (tmp > (IN_CLASSB_HOST - 0x0200));
 110         return htonl((LINKLOCAL_ADDR + 0x0100) + tmp);
 111 }
 112
 113 /**
 114  * Broadcast an ARP packet.
 115  */
 116 static void arp(
 117         /* int op, - always ARPOP_REQUEST */
 118         /* const struct ether_addr *source_eth, - always &eth_addr */
 119                                         struct in_addr source_ip,
 120         const struct ether_addr *target_eth, struct in_addr target_ip)
 121 {
 122         enum { op = ARPOP_REQUEST };
 123 #define source_eth (&eth_addr)
 124
 125         struct arp_packet p;
 126         memset(&p, 0, sizeof(p));
 127
 128         // ether header
 129         p.eth.ether_type = htons(ETHERTYPE_ARP);
 130         memcpy(p.eth.ether_shost, source_eth, ETH_ALEN);
 131         memset(p.eth.ether_dhost, 0xff, ETH_ALEN);
 132
 133         // arp request
 134         p.arp.arp_hrd = htons(ARPHRD_ETHER);
 135         p.arp.arp_pro = htons(ETHERTYPE_IP);
 136         p.arp.arp_hln = ETH_ALEN;
 137         p.arp.arp_pln = 4;
 138         p.arp.arp_op = htons(op);
 139         memcpy(&p.arp.arp_sha, source_eth, ETH_ALEN);
 140         memcpy(&p.arp.arp_spa, &source_ip, sizeof(p.arp.arp_spa));
 141         memcpy(&p.arp.arp_tha, target_eth, ETH_ALEN);
 142         memcpy(&p.arp.arp_tpa, &target_ip, sizeof(p.arp.arp_tpa));
 143
 144         // send it
 145         // Even though sock_fd is already bound to saddr, just send()
 146         // won't work, because "socket is not connected"
 147         // (and connect() won't fix that, "operation not supported").
 148         // Thus we sendto() to saddr. I wonder which sockaddr
 149         // (from bind() or from sendto()?) kernel actually uses
 150         // to determine iface to emit the packet from...
 151         xsendto(sock_fd, &p, sizeof(p), &saddr, sizeof(saddr));
 152 #undef source_eth
 153 }
 154
 155 /**
 156  * Run a script.
 157  * argv[0]:intf argv[1]:script_name argv[2]:junk argv[3]:NULL
 158  */
 159 static int run(char *argv[3], const char *param, struct in_addr *ip)
 160 {
 161         int status;
 162         char *addr = addr; /* for gcc */
 163         const char *fmt = "%s %s %s" + 3;
 164
 165         argv[2] = (char*)param;
 166
 167         VDBG("%s run %s %s\n", argv[0], argv[1], argv[2]);
 168
 169         if (ip) {
 170                 addr = inet_ntoa(*ip);
 171                 xsetenv("ip", addr);
 172                 fmt -= 3;
 173         }
 174         bb_info_msg(fmt, argv[2], argv[0], addr);
 175
 176         status = spawn_and_wait(argv + 1);
 177         if (status < 0) {
 178                 bb_perror_msg("%s %s %s" + 3, argv[2], argv[0]);
 179                 return -errno;
 180         }
 181         if (status != 0)
 182                 bb_error_msg("script %s %s failed, exitcode=%d", argv[1], argv[2], status & 0xff);
 183         return status;
 184 }
 185
 186 /**
 187  * Return milliseconds of random delay, up to "secs" seconds.
 188  */
 189 static ALWAYS_INLINE unsigned random_delay_ms(unsigned secs)
 190 {
 191         return rand() % (secs * 1000);
 192 }
 193
 194 /**
 195  * main program
 196  */
 197 int zcip_main(int argc, char **argv) MAIN_EXTERNALLY_VISIBLE;
 198 int zcip_main(int argc UNUSED_PARAM, char **argv)
 199 {
 200         int state;
 201         char *r_opt;
 202         unsigned opts;
 203
 204         // ugly trick, but I want these zeroed in one go
 205         struct {
 206                 const struct in_addr null_ip;
 207                 const struct ether_addr null_addr;
 208                 struct in_addr ip;
 209                 struct ifreq ifr;
 210                 int timeout_ms; /* must be signed */
 211                 unsigned conflicts;
 212                 unsigned nprobes;
 213                 unsigned nclaims;
 214                 int ready;
 215                 int verbose;
 216         } L;
 217 #define null_ip    (L.null_ip   )
 218 #define null_addr  (L.null_addr )
 219 #define ip         (L.ip        )
 220 #define ifr        (L.ifr       )
 221 #define timeout_ms (L.timeout_ms)
 222 #define conflicts  (L.conflicts )
 223 #define nprobes    (L.nprobes   )
 224 #define nclaims    (L.nclaims   )
 225 #define ready      (L.ready     )
 226 #define verbose    (L.verbose   )
 227
 228         memset(&L, 0, sizeof(L));
 229
 230 #define FOREGROUND (opts & 1)
 231 #define QUIT       (opts & 2)
 232         // parse commandline: prog [options] ifname script
 233         // exactly 2 args; -v accumulates and implies -f
 234         opt_complementary = "=2:vv:vf";
 235         opts = getopt32(argv, "fqr:v", &r_opt, &verbose);
 236 #if !BB_MMU
 237         // on NOMMU reexec early (or else we will rerun things twice)
 238         if (!FOREGROUND)
 239                 bb_daemonize_or_rexec(0 /*was: DAEMON_CHDIR_ROOT*/, argv);
 240 #endif
 241         // open an ARP socket
 242         // (need to do it before openlog to prevent openlog from taking
 243         // fd 3 (sock_fd==3))
 244         xmove_fd(xsocket(AF_PACKET, SOCK_PACKET, htons(ETH_P_ARP)), sock_fd);
 245         if (!FOREGROUND) {
 246                 // do it before all bb_xx_msg calls
 247                 openlog(applet_name, 0, LOG_DAEMON);
 248                 logmode |= LOGMODE_SYSLOG;
 249         }
 250         if (opts & 4) { // -r n.n.n.n
 251                 if (inet_aton(r_opt, &ip) == 0
 252                  || (ntohl(ip.s_addr) & IN_CLASSB_NET) != LINKLOCAL_ADDR
 253                 ) {
 254                         bb_error_msg_and_die("invalid link address");
 255                 }
 256         }
 257         argv += optind - 1;
 258
 259         /* Now: argv[0]:junk argv[1]:intf argv[2]:script argv[3]:NULL */
 260         /* We need to make space for script argument: */
 261         argv[0] = argv[1];
 262         argv[1] = argv[2];
 263         /* Now: argv[0]:intf argv[1]:script argv[2]:junk argv[3]:NULL */
 264 #define argv_intf (argv[0])
 265
 266         xsetenv("interface", argv_intf);
 267
 268         // initialize the interface (modprobe, ifup, etc)
 269         if (run(argv, "init", NULL))
 270                 return EXIT_FAILURE;
 271
 272         // initialize saddr
 273         // saddr is: { u16 sa_family; u8 sa_data[14]; }
 274         //memset(&saddr, 0, sizeof(saddr));
 275         //TODO: are we leaving sa_family == 0 (AF_UNSPEC)?!
 276         safe_strncpy(saddr.sa_data, argv_intf, sizeof(saddr.sa_data));
 277
 278         // bind to the interface's ARP socket
 279         xbind(sock_fd, &saddr, sizeof(saddr));
 280
 281         // get the interface's ethernet address
 282         //memset(&ifr, 0, sizeof(ifr));
 283         strncpy_IFNAMSIZ(ifr.ifr_name, argv_intf);
 284         xioctl(sock_fd, SIOCGIFHWADDR, &ifr);
 285         memcpy(&eth_addr, &ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 286
 287         // start with some stable ip address, either a function of
 288         // the hardware address or else the last address we used.
 289         // we are taking low-order four bytes, as top-order ones
 290         // aren't random enough.
 291         // NOTE: the sequence of addresses we try changes only
 292         // depending on when we detect conflicts.
 293         {
 294                 uint32_t t;
 295                 move_from_unaligned32(t, ((char *)&eth_addr + 2));
 296                 srand(t);
 297         }
 298         if (ip.s_addr == 0)
 299                 ip.s_addr = pick();
 300
 301         // FIXME cases to handle:
 302         //  - zcip already running!
 303         //  - link already has local address... just defend/update
 304
 305         // daemonize now; don't delay system startup
 306         if (!FOREGROUND) {
 307 #if BB_MMU
 308                 bb_daemonize(0 /*was: DAEMON_CHDIR_ROOT*/);
 309 #endif
 310                 bb_info_msg("start, interface %s", argv_intf);
 311         }
 312
 313         // run the dynamic address negotiation protocol,
 314         // restarting after address conflicts:
 315         //  - start with some address we want to try
 316         //  - short random delay
 317         //  - arp probes to see if another host uses it
 318         //  - arp announcements that we're claiming it
 319         //  - use it
 320         //  - defend it, within limits
 321         // exit if:
 322         // - address is successfully obtained and -q was given:
 323         //   run "<script> config", then exit with exitcode 0
 324         // - poll error (when does this happen?)
 325         // - read error (when does this happen?)
 326         // - sendto error (in arp()) (when does this happen?)
 327         // - revents & POLLERR (link down). run "<script> deconfig" first
 328         state = PROBE;
 329         while (1) {
 330                 struct pollfd fds[1];
 331                 unsigned deadline_us;
 332                 struct arp_packet p;
 333                 int source_ip_conflict;
 334                 int target_ip_conflict;
 335
 336                 fds[0].fd = sock_fd;
 337                 fds[0].events = POLLIN;
 338                 fds[0].revents = 0;
 339
 340                 // poll, being ready to adjust current timeout
 341                 if (!timeout_ms) {
 342                         timeout_ms = random_delay_ms(PROBE_WAIT);
 343                         // FIXME setsockopt(sock_fd, SO_ATTACH_FILTER, ...) to
 344                         // make the kernel filter out all packets except
 345                         // ones we'd care about.
 346                 }
 347                 // set deadline_us to the point in time when we timeout
 348                 deadline_us = MONOTONIC_US() + timeout_ms * 1000;
 349
 350                 VDBG("...wait %d %s nprobes=%u, nclaims=%u\n",
 351                                 timeout_ms, argv_intf, nprobes, nclaims);
 352
 353                 switch (safe_poll(fds, 1, timeout_ms)) {
 354
 355                 default:
 356                         //bb_perror_msg("poll"); - done in safe_poll
 357                         return EXIT_FAILURE;
 358
 359                 // timeout
 360                 case 0:
 361                         VDBG("state = %d\n", state);
 362                         switch (state) {
 363                         case PROBE:
 364                                 // timeouts in the PROBE state mean no conflicting ARP packets
 365                                 // have been received, so we can progress through the states
 366                                 if (nprobes < PROBE_NUM) {
 367                                         nprobes++;
 368                                         VDBG("probe/%u %s@%s\n",
 369                                                         nprobes, argv_intf, inet_ntoa(ip));
 370                                         arp(/* ARPOP_REQUEST, */
 371                                                         /* &eth_addr, */ null_ip,
 372                                                         &null_addr, ip);
 373                                         timeout_ms = PROBE_MIN * 1000;
 374                                         timeout_ms += random_delay_ms(PROBE_MAX - PROBE_MIN);
 375                                 }
 376                                 else {
 377                                         // Switch to announce state.
 378                                         state = ANNOUNCE;
 379                                         nclaims = 0;
 380                                         VDBG("announce/%u %s@%s\n",
 381                                                         nclaims, argv_intf, inet_ntoa(ip));
 382                                         arp(/* ARPOP_REQUEST, */
 383                                                         /* &eth_addr, */ ip,
 384                                                         &eth_addr, ip);
 385                                         timeout_ms = ANNOUNCE_INTERVAL * 1000;
 386                                 }
 387                                 break;
 388                         case RATE_LIMIT_PROBE:
 389                                 // timeouts in the RATE_LIMIT_PROBE state mean no conflicting ARP packets
 390                                 // have been received, so we can move immediately to the announce state
 391                                 state = ANNOUNCE;
 392                                 nclaims = 0;
 393                                 VDBG("announce/%u %s@%s\n",
 394                                                 nclaims, argv_intf, inet_ntoa(ip));
 395                                 arp(/* ARPOP_REQUEST, */
 396                                                 /* &eth_addr, */ ip,
 397                                                 &eth_addr, ip);
 398                                 timeout_ms = ANNOUNCE_INTERVAL * 1000;
 399                                 break;
 400                         case ANNOUNCE:
 401                                 // timeouts in the ANNOUNCE state mean no conflicting ARP packets
 402                                 // have been received, so we can progress through the states
 403                                 if (nclaims < ANNOUNCE_NUM) {
 404                                         nclaims++;
 405                                         VDBG("announce/%u %s@%s\n",
 406                                                         nclaims, argv_intf, inet_ntoa(ip));
 407                                         arp(/* ARPOP_REQUEST, */
 408                                                         /* &eth_addr, */ ip,
 409                                                         &eth_addr, ip);
 410                                         timeout_ms = ANNOUNCE_INTERVAL * 1000;
 411                                 }
 412                                 else {
 413                                         // Switch to monitor state.
 414                                         state = MONITOR;
 415                                         // link is ok to use earlier
 416                                         // FIXME update filters
 417                                         run(argv, "config", &ip);
 418                                         ready = 1;
 419                                         conflicts = 0;
 420                                         timeout_ms = -1; // Never timeout in the monitor state.
 421
 422                                         // NOTE: all other exit paths
 423                                         // should deconfig ...
 424                                         if (QUIT)
 425                                                 return EXIT_SUCCESS;
 426                                 }
 427                                 break;
 428                         case DEFEND:
 429                                 // We won!  No ARP replies, so just go back to monitor.
 430                                 state = MONITOR;
 431                                 timeout_ms = -1;
 432                                 conflicts = 0;
 433                                 break;
 434                         default:
 435                                 // Invalid, should never happen.  Restart the whole protocol.
 436                                 state = PROBE;
 437                                 ip.s_addr = pick();
 438                                 timeout_ms = 0;
 439                                 nprobes = 0;
 440                                 nclaims = 0;
 441                                 break;
 442                         } // switch (state)
 443                         break; // case 0 (timeout)
 444
 445                 // packets arriving, or link went down
 446                 case 1:
 447                         // We need to adjust the timeout in case we didn't receive
 448                         // a conflicting packet.
 449                         if (timeout_ms > 0) {
 450                                 unsigned diff = deadline_us - MONOTONIC_US();
 451                                 if ((int)(diff) < 0) {
 452                                         // Current time is greater than the expected timeout time.
 453                                         // Should never happen.
 454                                         VDBG("missed an expected timeout\n");
 455                                         timeout_ms = 0;
 456                                 } else {
 457                                         VDBG("adjusting timeout\n");
 458                                         timeout_ms = (diff / 1000) | 1; /* never 0 */
 459                                 }
 460                         }
 461
 462                         if ((fds[0].revents & POLLIN) == 0) {
 463                                 if (fds[0].revents & POLLERR) {
 464                                         // FIXME: links routinely go down;
 465                                         // this shouldn't necessarily exit.
 466                                         bb_error_msg("iface %s is down", argv_intf);
 467                                         if (ready) {
 468                                                 run(argv, "deconfig", &ip);
 469                                         }
 470                                         return EXIT_FAILURE;
 471                                 }
 472                                 continue;
 473                         }
 474
 475                         // read ARP packet
 476                         if (safe_read(sock_fd, &p, sizeof(p)) < 0) {
 477                                 bb_perror_msg_and_die(bb_msg_read_error);
 478                         }
 479                         if (p.eth.ether_type != htons(ETHERTYPE_ARP))
 480                                 continue;
 481 #ifdef DEBUG
 482                         {
 483                                 struct ether_addr *sha = (struct ether_addr *) p.arp.arp_sha;
 484                                 struct ether_addr *tha = (struct ether_addr *) p.arp.arp_tha;
 485                                 struct in_addr *spa = (struct in_addr *) p.arp.arp_spa;
 486                                 struct in_addr *tpa = (struct in_addr *) p.arp.arp_tpa;
 487                                 VDBG("%s recv arp type=%d, op=%d,\n",
 488                                         argv_intf, ntohs(p.eth.ether_type),
 489                                         ntohs(p.arp.arp_op));
 490                                 VDBG("\tsource=%s %s\n",
 491                                         ether_ntoa(sha),
 492                                         inet_ntoa(*spa));
 493                                 VDBG("\ttarget=%s %s\n",
 494                                         ether_ntoa(tha),
 495                                         inet_ntoa(*tpa));
 496                         }
 497 #endif
 498                         if (p.arp.arp_op != htons(ARPOP_REQUEST)
 499                          && p.arp.arp_op != htons(ARPOP_REPLY))
 500                                 continue;
 501
 502                         source_ip_conflict = 0;
 503                         target_ip_conflict = 0;
 504
 505                         if (memcmp(p.arp.arp_spa, &ip.s_addr, sizeof(struct in_addr)) == 0
 506                          && memcmp(&p.arp.arp_sha, &eth_addr, ETH_ALEN) != 0
 507                         ) {
 508                                 source_ip_conflict = 1;
 509                         }
 510                         if (p.arp.arp_op == htons(ARPOP_REQUEST)
 511                          && memcmp(p.arp.arp_tpa, &ip.s_addr, sizeof(struct in_addr)) == 0
 512                          && memcmp(&p.arp.arp_tha, &eth_addr, ETH_ALEN) != 0
 513                         ) {
 514                                 target_ip_conflict = 1;
 515                         }
 516
 517                         VDBG("state = %d, source ip conflict = %d, target ip conflict = %d\n",
 518                                 state, source_ip_conflict, target_ip_conflict);
 519                         switch (state) {
 520                         case PROBE:
 521                         case ANNOUNCE:
 522                                 // When probing or announcing, check for source IP conflicts
 523                                 // and other hosts doing ARP probes (target IP conflicts).
 524                                 if (source_ip_conflict || target_ip_conflict) {
 525                                         conflicts++;
 526                                         if (conflicts >= MAX_CONFLICTS) {
 527                                                 VDBG("%s ratelimit\n", argv_intf);
 528                                                 timeout_ms = RATE_LIMIT_INTERVAL * 1000;
 529                                                 state = RATE_LIMIT_PROBE;
 530                                         }
 531
 532                                         // restart the whole protocol
 533                                         ip.s_addr = pick();
 534                                         timeout_ms = 0;
 535                                         nprobes = 0;
 536                                         nclaims = 0;
 537                                 }
 538                                 break;
 539                         case MONITOR:
 540                                 // If a conflict, we try to defend with a single ARP probe.
 541                                 if (source_ip_conflict) {
 542                                         VDBG("monitor conflict -- defending\n");
 543                                         state = DEFEND;
 544                                         timeout_ms = DEFEND_INTERVAL * 1000;
 545                                         arp(/* ARPOP_REQUEST, */
 546                                                 /* &eth_addr, */ ip,
 547                                                 &eth_addr, ip);
 548                                 }
 549                                 break;
 550                         case DEFEND:
 551                                 // Well, we tried.  Start over (on conflict).
 552                                 if (source_ip_conflict) {
 553                                         state = PROBE;
 554                                         VDBG("defend conflict -- starting over\n");
 555                                         ready = 0;
 556                                         run(argv, "deconfig", &ip);
 557
 558                                         // restart the whole protocol
 559                                         ip.s_addr = pick();
 560                                         timeout_ms = 0;
 561                                         nprobes = 0;
 562                                         nclaims = 0;
 563                                 }
 564                                 break;
 565                         default:
 566                                 // Invalid, should never happen.  Restart the whole protocol.
 567                                 VDBG("invalid state -- starting over\n");
 568                                 state = PROBE;
 569                                 ip.s_addr = pick();
 570                                 timeout_ms = 0;
 571                                 nprobes = 0;
 572                                 nclaims = 0;
 573                                 break;
 574                         } // switch state
 575                         break; // case 1 (packets arriving)
 576                 } // switch poll
 577         } // while (1)
 578 #undef argv_intf
 579 }